CN114578249A - Lithium battery health state estimation method based on stability characteristics and AS-TCN model - Google Patents

Abstract

The invention relates to an improvement of a lithium battery health state evaluation method, in particular to a lithium battery health state evaluation method based on a stable characteristic and an AS-TCN model, which can realize accurate and efficient SOH monitoring and comprises the following steps: step S1: selecting a lithium battery to be tested to perform a charge and discharge experiment, wherein technical parameters needing to be collected in the experiment comprise a voltage value, a temperature value, a collection time point and a battery capacity value of a corresponding period under a constant-current charging working condition; step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each group of cyclic charge-discharge periods according to the same time node; step S3: clustering the acquired data through a dynamic time warping gravity center average algorithm, wherein the optimal comparison of the two sequences is determined under the condition that the dynamic time warping distance algorithm is staggered with respect to time, and then dividing the clustered voltage and temperature sequences into a training data set and a testing data set for model prediction; step S4: and constructing an attention separable time convolution network model.

Description

Lithium battery health state estimation method based on stability characteristics and AS-TCN model

Technical Field

The invention relates to improvement of a lithium battery health state evaluation method, in particular to a lithium battery health state evaluation method based on a stability characteristic and an AS-TCN model.

Background

Lithium ion batteries have been widely used in human life due to their characteristics of fast charging speed, low self-discharge, long service life, etc., and nowadays, application sites of lithium batteries have been expanded to larger application sites (such as new energy vehicles, unmanned aerial vehicles, and satellites) through portable electronic devices (such as mobile phones, cameras, and notebook computers), and other larger industrial devices and electric energy storage have been kept away. The occurrence of the phenomenon of total capacity reduction and internal resistance increase of the lithium ion battery in the process of recycling indicates that the battery has the characteristic of energy storage capacity decline, and the characteristic is generally called battery aging. The State of health (SOH) is a standard for evaluating the aging degree of the battery, and essentially reflects the aging and damage degree of the lithium battery. The aging of the battery is a process of the degradation of each function of the battery, and the aging is a complex nonlinear process and involves a very wide range of factors. Generally, when the actual capacity of the battery is reduced to 70% -80% of the standard capacity, the service life of the battery is over, and if the battery is used continuously, the operation performance of the whole system can be affected, so that disasters can occur.

For the current research situation, the equivalent circuit model is formed by electronic devices to simulate the battery characteristics, and has lower difficulty than the electrochemical model, high realizability and high dynamic response, but the essence of the equivalent circuit is approximate processing, and the parameter deviation of part of the model can cause larger prediction error. The electrochemical model can accurately estimate the SOH of the battery mainly by researching the electrochemical reaction process in the battery, but the health estimation based on the electrochemical model is difficult to model and has high practical application difficulty. The method based on the mechanism model is difficult to realize and has no universal applicability, so that more researches are carried out on a data driving model at present, the data driving model is classified into an empirical model, and the aim of accurate prediction is fulfilled by training a large amount of data.

Aiming at the current research situation of data driving, a data driving model consists of two parts, namely a feature extraction part and a model estimation part. For the feature extraction part, the following can be found from the current research situation:

1. different feature extraction methods have different degrees of association obtained by batteries under different working conditions, and a feature extraction method which is high in precision and suitable for all batteries does not exist at present.

2. The current feature extraction-based method only analyzes the importance of features from the relevance of feature factors and capacity degradation trends, neglects the influence of the feature with high relevance on a prediction result essentially improved from a model, and has the limitation of depending on the whole charging and discharging process.

3. There is a great instability in the discharge process,

for the model estimation part, it can be found from the existing research status:

the models that can be improved by the SOH estimation method based on the health characteristics are mostly focused on traditional algorithms or models that have proven to be excellent in the field of battery performance, but there is still much room for improvement.

Therefore, it is important to devise a method that enables accurate and efficient monitoring of SOH to ensure that the battery is replaced before it fails.

Disclosure of Invention

In order to solve the problem, the invention provides a lithium battery health state estimation method based on a stability characteristic and an AS-TCN model, and SOH can be accurately and efficiently monitored.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows: a lithium battery health state estimation method based on a stable characteristic and an AS-TCN model comprises the following steps:

step S1: selecting a lithium battery to be tested to perform a charge and discharge experiment, wherein technical parameters needing to be collected in the experiment comprise a voltage value, a temperature value, a collection time point and a battery capacity value of a corresponding period under a constant-current charging working condition;

step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each group of cyclic charge-discharge periods according to the same time node;

step S3: clustering the acquired data through a dynamic time warping gravity center average algorithm, wherein the optimal comparison of the two sequences is determined under the condition that the dynamic time warping distance algorithm is staggered with respect to time, and then dividing the clustered voltage and temperature sequences into a training data set and a testing data set for model prediction;

step S4: constructing an attention separable time convolution network model, which comprises a depth separable convolution improved structure and a convolution attention mechanism model;

step S5: and estimating the health state of the lithium battery based on the established AS-TCN model.

Preferably, the curve fitting and data preprocessing for the data in step S2 specifically include: and respectively fitting corresponding time sequences to the collected two technical parameter sequence groups of the voltage and the temperature of the lithium battery according to the same collection time node, wherein the result is a numerical value change time sequence of the temperature and the voltage on the same charge time node along with the aging of the battery under different charge-discharge cycles.

Preferably, the clustering the acquired data by the dynamic time warping center-of-gravity averaging algorithm in step S3 specifically includes:

step S31: randomly selecting a sequence from the current sequence set S as an average sequence;

step S32: calculating the DTW distance of the selected average sequence and each sequence in the sequence set, and matching the coordinates of the average sequence with the coordinates of other sequences in the sequence set;

step S33: according to the DTW distance algorithm, it is assumed that the sequence a is present (a)₁,a₂,...,a_m)，B＝(b₁,b₂,...,b_n) And the DTW distance between A and B is recorded as D (A)_i,B_j) Indicating a DTW distance value between the ith time point of the sequence A and the jth time point of the sequence B;

step S34: comparing with the first calculated DTW distance, if the distance is the same, executing step S36, otherwise executing step S35;

step S35: judging whether the square sum of the DTW distance is reduced, if so, executing a step S36, and if not, obtaining the current average sequence as a result;

step S36: and updating each coordinate of the average sequence to be the coordinate mean value matched with the coordinate in the sequence set, and then returning to the step S32 to continue the iteration.

Preferably, the step S33, in which the DTW distance algorithm determines the optimal alignment of two sequences under the condition of time misalignment, specifically includes:

(1) definition of boundary conditions: given w₁＝(1,1)，w_k(m, n), the start and end points of the curved path must be the first and last points of the aligned time series;

(2) definition of monotonicity conditions: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≥0，y_k-y_k-1≥0；

(3) Definition of step size condition: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≤1，y_k-y _k-11 or less, and the basic step length condition is w_k-w_k-1∈{(1,1),(1,0),(0,1)}；

(4) Cumulative distance d_wThe calculation formula of (A, B):

where is the set of all possible paths;

(5) the minimum value of the cumulative distance is calculated by the formula:

(6) the cumulative distance matrix is D, D (i, j) is the set of all elements of the distance matrix D, and the calculation formula is:

D(i,j)＝C_ij+min{C_i-1,j-1,C_i-1,j,C_i,j-1},i∈[1:m],j∈[1:n] (5)。

preferably, in step S4, the depth separable convolution improvement structure includes:

(1) the parameter calculation formula after the time convolution network optimization based on the depth separable convolution is as follows: sxk × 1+ sx 1 × 1 × n;

wherein, nxS is the dimension of input, S is the length of the sequence, and n is the number of the sequence; k × 1 represents the size of the convolution kernel;

(2) the time convolution network input dimensionality based on the depth separable convolution is as follows: a is multiplied by b;

wherein a and b represent the number and length of input sequences, the value ranges of a and b are increased, a is more than or equal to 1, and b is more than or equal to 1;

(3) the specific location of the depth separable convolution in the time convolutional network structure: the Depthwise convolution of the deep separable convolution exists at the input position of the network and is used for automatically extracting the features, and the Pointwise convolution is used for recombining the extracted features before the full connection layer;

(4) selecting an activation function as Leaky Relu;

preferably, in step S4, the convolution attention mechanism model includes:

(1) the channel attention calculation formula is:

where Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W₀∈R^C/r×1，W₁∈R^1×C/rSigma represents Sigmoid activation operation, F represents a feature graph, C represents the third dimension of the feature graph, and r is the reduction rate;

(2) the spatial attention characteristic calculation formula is as follows:

wherein Ms represents the spatial attention feature, 7 × 7 represents the convolution kernel size, and s represents the feature map second dimension size;

(3) the final generated feature F' is calculated as follows:

(4) the specific positions of the convolution attention mechanism in the AS-TCN model are AS follows: after batch normalization, before the Leaky Relu activation function.

Preferably, in step S4, each residual module of the AS-TCN model is composed of two sub-modules, the middle of each residual module is connected by a residual structure, and the sub-modules have the following structures: mixed dilation-causal convolution, batch normalization, convolution attention mechanism, LeakyReLu activation function, and dropout; the integral structure is as follows: an input layer, a Depthwise convolution operation, a residual block, a Ponitwise convolution operation, and an output layer.

Preferably, in step S5, estimating the state of health of the lithium battery based on the established AS-TCN model specifically includes the following steps:

step S51: determining technical parameters needing to be acquired: current, voltage, time node and lithium battery capacity;

step S52: data preprocessing: the method comprises the steps of denoising data and supplementing missing values, and respectively fitting corresponding time sequences to two collected technical parameter sequence groups of the voltage and the temperature of the lithium battery according to the same collection time node;

step S53: data input: clustering the preprocessed voltage and temperature data through a DBA algorithm to remove data space coupling, and inputting current and voltage data of any segment;

step S54: training a model: the data is divided into 7:3 dividing the model into training data and testing data, and adjusting the hyper-parameters in the model by a control variable method;

step S55: selection of regression evaluation criteria: selecting a mean square error and a root mean square error as indexes for evaluating the accuracy of the model;

step S56: and (3) model verification: and (3) verifying the model by using 30% of test data, and proving the quality of the model by regression evaluation indexes.

The lithium battery health state estimation method based on the stability characteristics and the AS-TCN model can achieve the following beneficial effects:

(1) the mode of automatic feature extraction replaces manual feature extraction, uses the separable convolution feature extraction layer of degree of depth to extract the feature automatically, and the feature that obtains like this is more stable and has universal applicability more, regardless of the type of battery.

(2) The prediction can be completed only by partial data, the key point is that the correlation analysis is not needed to judge the quality of the characteristics, the original data and the prediction model are directly adjusted from the prediction result instead of the adjustment of manually extracted characteristics, and the reliability is higher.

(3) The voltage and temperature values of the battery at different aging degrees are represented by taking the acquisition time as a scale, a large amount of redundant data are found, the voltage and temperature data have a stable stage from the graph, the data of the stage have high similarity in space, the stable stage is taken as a main part, and the voltage and temperature sequence of the stable stage is clustered by a DBA algorithm, namely the original trend of the data is kept, and the calculation burden of a prediction model is also reduced.

(4) An AS-TCN model with an improved structure is built, the model is based on a light model TCN, the calculation amount is small, the precision is high, the defects of the original TCN are overcome for a multi-sequence prediction model, and the model is easier to use in practice.

Drawings

FIG. 1 is a flow chart of a lithium battery state of health estimation method of the present invention;

FIG. 2 is a time series of numerical changes in temperature and voltage at the same charge time node as the battery ages;

FIG. 3 is the result of DBA clustering of partial voltage and temperature data;

FIG. 4 is a diagram of the model structure of the AS-TCN and the techniques used by each layer of the network;

fig. 5 is a diagram illustrating the results of off-line estimation of unknown battery capacity.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.

A lithium battery SOH estimation method based on stable characteristic and AS-TCN neural network, said method is based on the stable degradation trend that the capacity exists in the charging process, reflect to the technical parameter that is voltage and temperature have stable degradation trend along with aging of the battery, regard this trend AS the foundation to judge the residual capacity of the battery directly, therefore said method has better stability, can separate the automatic extraction characteristic of the convolution layer through the depth, can avoid the error problem that the manual extraction health characteristic exists, have universal serviceability too; then clustering the health characteristics by using a DBA algorithm and taking time nodes as a standard so as to achieve the purposes of compressing data and not changing the overall trend; and finally, establishing an AS-TCN estimation model, in the process of establishing the model, taking the TCN AS an original model, realizing automatic feature extraction on input multidimensional health features through Depthwise convolution in a deep separable convolution technology, solving the problem of excessive neuron death caused by small feature variation by using a Leaky Relu activation function in the process, improving the overall prediction precision of the model through a convolution attention mechanism, capturing fine features, inhibiting redundant features which cannot be removed manually, and finally collecting all the features in a full connection layer for final prediction through Pointwise convolution.

In specific implementation, the process is shown in fig. 1, and is specifically implemented according to the following steps:

step S1: selecting a lithium battery to be tested to perform a charge and discharge experiment, wherein technical parameters needing to be collected in the experiment comprise a voltage value, a temperature value, a collection time point and a battery capacity value of a corresponding period under a constant-current charging working condition;

step S2: and respectively fitting corresponding time sequences to the collected two technical parameter sequence groups of the voltage and the temperature of the lithium battery according to the same collection time node, wherein the result is a numerical value change time sequence of the temperature and the voltage on the same charging time node along with the aging of the battery under different charging and discharging periods. As shown in fig. 2 in particular, it can be seen that there is a stable phase of voltage and temperature data, and the data have a high similarity in space;

step S3: clustering spatial height similarity data existing in voltage and temperature by using a Dynamic Time Warping Barycenter Averaging (DBA), specifically comprising the following steps:

step S31, randomly selecting a sequence from the current sequence set S as an average sequence;

and step S32, calculating the DTW distance of the selected average sequence and each sequence in the sequence set, and matching the coordinates of the average sequence with the coordinates of other sequences in the sequence set.

Step S33: according to the DTW distance algorithm, assume that there is a sequence a ═ (a)₁,a₂,...,a_m)，B＝(b₁,b₂,...,b_n) And the DTW distance between A and B is recorded as D (A)_i,B_j) Indicating a DTW distance value between the ith time point of the sequence A and the jth time point of the sequence B;

step S34: comparing with the first calculated DTW distance, if the distance is the same, executing step S36, otherwise executing step S35;

step S35: judging whether the square sum of the DTW distance is reduced, if so, executing a step S36, and if not, obtaining the current average sequence as a result;

step S36: and updating each coordinate of the average sequence to be the coordinate mean value matched with the average sequence in the sequence set, and then returning to the step S32 to continue the iteration.

Step S33 in the DBA algorithm includes determining an optimal alignment of two sequences under a condition that a Dynamic Time Warping (DTW) distance algorithm is misaligned with respect to time, which specifically includes:

(1) definition of boundary conditions: given w₁＝(1,1)，w_k＝(mN), the starting and ending points of the curved path must be the first and last points of the aligned time series.

(2) Definition of monotonicity conditions: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≥0，y_k-y_k-1≧ 0, which makes the points in W monotonically spaced in time.

(3) Definition of step size condition: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≤1，y_k-y _k-11 or less, and the basic step length condition is w_k-w_k-1E { (1,1), (1,0), (0,1) }, which limits the long-distance jumps of the curved path when the sequences are aligned.

(4) Cumulative distance d_wThe calculation formula of (A, B):

wherein W ∈ W ═ W₁,w₂,...,w_k) Is the set of all possible paths

(5) The minimum value of the cumulative distance is calculated by the formula:

(6) the cumulative distance matrix is D, D (i, j) is the set of all elements of the distance matrix D, and the calculation formula is:

D(i,j)＝C_ij+min{C_i-1,j-1,C_i-1,j,C_i,j-1},i∈[1:m],j∈[1:n] (5)

and then dividing the clustered voltage and temperature sequences into a training data set and a testing data set for model prediction, wherein FIG. 3 shows the result of partial voltage and temperature data after DBA clustering, and the overall trend of the partial data can be found by the DBA algorithm.

Step S4: an Attention Separable Temporal Convolutional network (AS-TCN) model is built, the AS-TCN model comprises a depth Separable convolution improved structure and a convolution Attention mechanism model, the AS-TCN model structure and the technology used by each layer of the network are shown in FIG. 4, after original data are input, an initial feature extraction stage of Depthwise is firstly carried out, and the process of input value taking and calculation of the improved model in terms of the model structure comprises the following steps:

(1) the time convolution network input dimensionality based on the depth separable convolution is as follows: and a is multiplied by b, wherein a and b represent the number and the length of input sequences, the value ranges of a and b are increased, a is larger than or equal to 1, and b is larger than or equal to 1. In the study, the voltage and the temperature are selected as input sequences, namely, the value of a is 2.

(2) The parameter calculation formula after the time convolution network optimization based on the depth separable convolution is as follows: sxkx1 + sx1 × 1 × 1 × n, and the number of parameters of the ordinary convolution is sxkx1 × n, where nxs is the input dimension, S is the length of the sequence, and n is the number of the sequences; k × 1 represents the size of the convolution kernel. From the formula, it can be found that the parameters are relatively fewer when the input-output dimension n of the model is larger, so that the method has the advantage of small calculation amount for voltage and temperature data with larger data amount.

(3) The specific location of the depth separable convolution in the AS-TCN structure: the Depthwise convolution of the depth separable convolution exists at the input position of the network, so that the Depthwise convolution has two advantages, namely, the Depthwise convolution is used for automatically extracting the features so as to obtain stable features; and secondly, the TCN network is improved into a novel structure of multivariate sequence prediction. The poitwise convolution is used to recombine the extracted features before the full connected layer.

After the automatic extraction of the Depthwise characteristics, further optimizing the characteristics by a convolution attention mechanism model:

(1) attention of the channel:

in order to improve the attention capacity of the model to important information and pay attention to factors really influencing the prediction result, not only from the analysis of relevance. The lightweight convolution attention module is integrated in the model and is used as a method for improving the representation capability of the CNN network, so that the learning capability of the model to important features is simply and efficiently improved. The research converts the data from the three-dimensional image processing field into two dimensions for processing sequence data and embeds the data into a TCN structure. The lightweight convolution attention module consists of a channel attention mechanism and a space attention mechanism, is used for increasing the expression capacity of features, and has better effect compared with the attention mechanism which mostly focuses on only channels. The model firstly carries out average pooling operation and maximum pooling downsampling on the intermediate feature map in channel dimension, compresses global space information into a channel descriptor of a one-dimensional vector, puts the channel descriptor into a multilayer perceptron for adjustment, and finally generates a final channel attention feature map Mc through sigmoid activation operation

Where Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W₀∈R^C/r×1，W₁∈R^1×C/rσ represents Sigmoid activation operation, F represents a feature map, C represents the size of the third dimension of the feature map, and r is the reduction rate.

(2) And multiplying the channel attention feature map and the input feature map according to elements to generate the input features required by the spatial attention module. Then, in the spatial dimension, processing the feature graph processed in the channel dimension through the global average pooling and the global maximum pooling based on the channel, performing connection operation on the result based on the channel, performing convolution operation for dimensionality reduction, and finally outputting a spatial attention feature Ms through a sigmoid activation function, wherein a spatial attention feature calculation formula is as follows:

where Ms represents the spatial attention feature, 7 × 7 represents the convolution kernel size, and s represents the feature map second dimension size.

The Ms is multiplied with the input features of the module to obtain finally generated features F', and the features after denoising by the convolution attention mechanism can realize normalization before regression prediction by pointwise convolution.

Secondly, it should be noted that the specific positions of the convolution attention mechanism in the AS-TCN model are: after batch normalization, before the Leaky Relu activation function. The activation function selected as Leaky Relu is the combination of theory and experimental process, and solves the problem of neuron death in the actual estimation process by means of setting zero in a negative interval.

Step S5: and estimating the health state of the lithium battery based on the established AS-TCN model.

The topology of SOH estimation based on the AS-TCN model is AS follows: firstly, each residual error module of the AS-TCN model consists of two sub-modules, the middle of each residual error module is connected through a residual error structure, and the sub-modules have the following structures: mixed dilation-causal convolution, batch normalization, convolution attention mechanism, LeakyReLu activation function, and dropout. The integral structure is as follows: an input layer, a Depthwise convolution operation layer, a five-layer residual module, a Ponitwise convolution regulation layer and an output layer.

The method specifically comprises the following steps:

step S51: determining model input parameters and dimensions of the parameters: selecting voltage and temperature as input, and using corresponding capacity as an input label;

step S52: preprocessing before automatic feature extraction of data: the method comprises the steps of data denoising, missing value supplement and data normalization processing;

step S53: inputting a model: clustering the preprocessed voltage and temperature data through a DBA algorithm to remove data space coupling, and inputting current and voltage data of any segment;

step S54: training a model: dividing data into training data and testing data according to the ratio of 7:3, adjusting hyper-parameters in a model by a control variable method, selecting a loss function as a mean square error function, and selecting an adam optimization algorithm by an optimizer;

step S55: selection of regression evaluation criteria: selecting the average absolute error and the root-mean-square error as indexes for evaluating the accuracy of the model;

step S56: and (3) model verification: the model is verified by using 30% of test data, the quality of the model is proved by regression evaluation indexes, and fig. 5 shows the offline estimation result of the unknown battery capacity, wherein the root mean square error is 0.012, and the average absolute error is 0.007.

Finally, the above embodiments are only intended to illustrate the technical solutions of the present invention and not to limit the present invention, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions, and all of them should be covered by the claims of the present invention.

The above is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (8)

1. A lithium battery health state estimation method based on a stable characteristic and an AS-TCN model is characterized by comprising the following steps:

step S1: selecting a lithium battery to be tested to perform a charge and discharge experiment, wherein technical parameters needing to be collected in the experiment comprise a voltage value, a temperature value, a collection time point and a battery capacity value of a corresponding period under a constant-current charging working condition;

step S2: performing curve fitting and data preprocessing on the voltage and temperature data of each group of cyclic charge-discharge periods according to the same time node;

step S3: clustering the acquired data through a dynamic time warping gravity center average algorithm, wherein the optimal comparison of the two sequences is determined under the condition that the dynamic time warping distance algorithm is staggered with respect to time, and then dividing the clustered voltage and temperature sequences into a training data set and a testing data set for model prediction;

step S4: constructing an attention separable time convolution network model, which comprises a depth separable convolution improvement structure and a convolution attention mechanism model;

step S5: and estimating the health state of the lithium battery based on the established AS-TCN model.

2. The lithium battery state of health estimation method based on the stable characteristic and the AS-TCN model AS claimed in claim 1, wherein the curve fitting and data preprocessing for the data in the step S2 specifically comprises: and respectively fitting corresponding time sequences to the collected two technical parameter sequence groups of the voltage and the temperature of the lithium battery according to the same collection time node, wherein the result is a numerical value change time sequence of the temperature and the voltage on the same charge time node along with the aging of the battery under different charge-discharge cycles.

3. The lithium battery state of health estimation method based on stable features and AS-TCN model according to claim 1, characterized in that: the clustering of the collected data by the dynamic time warping gravity center average algorithm in step S3 specifically includes:

step S31: randomly selecting a sequence from the current sequence set S as an average sequence;

step S32: calculating the DTW distance of the selected average sequence and each sequence in the sequence set, and matching the coordinates of the average sequence with the coordinates of other sequences in the sequence set;

step S33: according to the DTW distance algorithm, it is assumed that the sequence a is present (a)₁,a₂,...,a_m)，B＝(b₁,b₂,...,b_n) And the DTW distance between A and B is recorded as D (A)_i,B_j) Indicating a DTW distance value between the ith time point of the sequence A and the jth time point of the sequence B;

step S34: comparing with the first calculated DTW distance, if the distance is the same, executing step S36, otherwise executing step S35;

step S35: judging whether the square sum of the DTW distance is reduced, if so, executing a step S36, and if not, obtaining the current average sequence as a result;

step S36: and updating each coordinate of the average sequence to be the coordinate mean value matched with the coordinate in the sequence set, and then returning to the step S32 to continue the iteration.

4. The lithium battery state of health estimation method based on stable characteristics and AS-TCN model according to claim 3, characterized in that: in step S33, the DTW distance algorithm determines the optimal alignment of the two sequences under the condition of time misalignment, and specifically includes:

(1) definition of boundary conditions: given w₁＝(1,1)，w_k(m, n), the start and end points of the curved path must be the first and last points of the aligned time series;

(2) definition of monotonicity conditions: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≥0，y_k-y_k-1≥0；

(3) Definition of step size condition: given w_k＝(x_k,y_k)，w_k-1＝(x_k-1,y_k-1) Wherein x is_k-x_k-1≤1，y_k-y_k-11 or less, the basic step length condition is w_k-w_k-1∈{(1,1),(1,0),(0,1)}；

(4) Cumulative distance d_wThe calculation formula of (A, B):

where is the set of all possible paths;

(5) the minimum value of the cumulative distance is calculated by the formula:

(6) the cumulative distance matrix is D, D (i, j) is the set of all elements of the distance matrix D, and the calculation formula is:

D(i,j)＝C_ij+min{C_i-1,j-1,C_i-1,j,C_i,j-1},i∈[1:m],j∈[1:n] (5)。

5. the lithium battery state of health estimation method based on stable features and AS-TCN model according to claim 1, characterized in that: in step S4, the depth separable convolution improvement structure includes:

(1) the parameter calculation formula after the time convolution network optimization based on the depth separable convolution is as follows: sxkx 1+ sx 1 × 1 × n;

wherein, nxS is the dimension of input, S is the length of the sequence, and n is the number of the sequence; k × 1 represents the size of the convolution kernel;

(2) the time convolution network input dimensionality based on the depth separable convolution is as follows: a is multiplied by b;

wherein a and b represent the number and length of input sequences, the value ranges of a and b are increased, a is more than or equal to 1, and b is more than or equal to 1;

(3) the specific location of the depth separable convolution in the time convolutional network structure: the Depthwise convolution of the deep separable convolution exists at the input position of the network and is used for automatically extracting the features, and the Pointwise convolution is used for recombining the extracted features before the full connection layer;

(4) the activation function is chosen to be Leaky Relu.

6. The lithium battery state of health estimation method based on stable features and AS-TCN model according to claim 1, characterized in that: in step S4, the convolution attention mechanism model includes:

(1) the channel attention calculation formula is:

where Mc represents the channel attention feature, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, W₀∈R^C/r×1，W₁∈R^1×C/rSigma represents Sigmoid activation operation, F represents a feature graph, C represents the third dimension of the feature graph, and r is the reduction rate;

(2) the spatial attention characteristic calculation formula is as follows:

wherein Ms represents the spatial attention feature, 7 × 7 represents the convolution kernel size, and s represents the feature map second dimension size;

(3) the final generated feature F' is calculated as follows:

(4) the specific positions of the convolution attention mechanism in the AS-TCN model are AS follows: after batch normalization, before the Leaky Relu activation function.

7. The lithium battery state of health estimation method based on stable characteristics and AS-TCN neural network of claim 1, characterized in that: in step S4, each residual module of the AS-TCN model is composed of two sub-modules, the middle of each residual module is connected by a residual structure, and the sub-modules have the following structures: mixed dilation-causal convolution, batch normalization, convolution attention mechanism, LeakyReLu activation function, and dropout; the integral structure is as follows: an input layer, a Depthwise convolution operation, a residual block, a Ponitwise convolution operation, and an output layer.

8. The lithium battery health state estimation method based on the stable characteristic and the AS-TCN neural network AS claimed in claim 1, wherein: in step S5, estimating the state of health of the lithium battery based on the established AS-TCN model, specifically including the steps of:

step S51: determining technical parameters needing to be acquired: current, voltage, time node and lithium battery capacity;

step S52: data preprocessing: the method comprises the steps of denoising data and supplementing missing values, and respectively fitting corresponding time sequences to two collected technical parameter sequence groups of the voltage and the temperature of the lithium battery according to the same collection time node;

step S53: data input: clustering the preprocessed voltage and temperature data through a DBA algorithm to remove data space coupling, and inputting current and voltage data of any segment;

step S54: training a model: the data is divided into 7:3 dividing the model into training data and testing data, and adjusting the hyper-parameters in the model by a control variable method;

step S55: selection of regression evaluation criteria: selecting a mean square error and a root mean square error as indexes for evaluating the accuracy of the model;

step S56: and (3) model verification: and (3) verifying the model by using 30% of test data, and proving the quality of the model by regression evaluation indexes.

Images